Adaptive Virtual Machine Provisioning in Elastic Multitier Cloud

Adaptive Virtual Machine Provisioning in Elastic Multitier Cloud Platforms Fan Zhang, Junwei Cao, Hong Cai James J. Mulcahy, Cheng Wu Tsinghua University, IBM, FAU 2011. 07. 28

Outline 1 Introduction & Related Works 2 System Architecture Overview 3 Virtualized Resource Scheduling 4 Experimental Studies Department of Automation, Tsinghua University

1. 1 Introduction (Background) n Virtualized Cloud Platform n n n S(Software)aa. S P(Platform)aa. S I(Infrastructure)aa. S Virtual Machines Virtual Clusters Advantages: (1)Creating/Destroying VM (2)Data/Processing locality (3)Service migration Department of Automation, Tsinghua University

1. 1 Introduction (Motivation) Operational Cost Users P er VM • How many VMro topuse? usage Use less VMs Response Time Workload estim Vendors ation • How. Rmany es VM to provide? p. Time estimation Use more VMs Department of Automation, Tsinghua University

1. 1 Introduction (Importance) User Vendor • Service level agreement • Lowing cost • High utilization • More customers Department of Automation, Tsinghua University

1. 2 Related Works n n n n n L. Slothouber. A model of web server performance. In Proceedings of the 5 th International World Wide Web Conference (WWW). Paris, France. 1996. J. Chase, and R. Doyle. Balance of power: Energy management for server clusters. In Proceedings of the 8 th Workshop on Hot Topics in Operating Systems (Hot. OS-VIII). Elmau, Germany. 2001. B. Urgaonkar, and P. Shenoy, Cataclysm: Handling extreme overloads in internet services. In Proceedings of the 23 rd Annual ACM SIGACT-SIGOPS Symposium on Principles of • None of the previous work considers the cost Distributed Computing (PODC’ 04). St. John’s, Newfoundland, Canada. 2004. of using VMs. R. Levy, J. Nagarajarao, G. Pacifici, M. Spreitzer, A. Tantawi, and A. Youssef. Performance management for cluster based web services. In IFIP/IEEE 8 th International Symposium on Integrated Network Management. Vol. 246, pp. 247– 261. 2003. D. Menasce, Web server software architectures. IEEE Internet Computing. Vol. 7 no. 6, 2003. • None of the previous D. Villela , P. Pradhan, and D. Rubenstein. Provisioning serverswork in the considers application tier for ecommerce systems. ACM Transactions on Internet Technology. VMs. (TOIT). Vol. 7, no. 1, 2007. providing large/small S. Ranjan, J. Rolia, H. FU, and E. Knightly. Qo. S-driven servermigration for internet data centers. In Proceedings of the 10 th International Workshop on Quality of Service(IWQo. S), Miami, FL. 2002. • Yaksha: Most of the previous work calculate A. Kamra, V. Misra, E. M. Nahum, a self-tuning controller for managing theresponse performance of 3 -tiered Web sites, time In Proceedings of thebased 12 th International Workshop on use estimation on simulation. We Quality of Service(IWQo. S), Passau, Germany, 2004. prediction, which is easier. mathematical B. Urgaonkar, G. Pacifici, P. Shenoy, M. Spreitzer, and A. Tantawi. Analytic Modeling of Multitier Internet Services and its Applications. ACM Transactions on the Web (TWEB 2007), Vol. 1, No. 1, pp. 1 -35, May 2007. We differentiate our work from the following three aspects. Cost considering Various VM Mathematical prediction Department of Automation, Tsinghua University

2. System Architecture Overview • Small VM • 1 CPU, 1 GB M. • • Virtual CPU Virtual Memory Virtual Machines Virtual Clusters Virtual Everything • Large VM • 2 CPU, 2 GB M. Department of Automation, Tsinghua University

2. System Architecture Overview j(i) (AARj) =AARj-1, j + AARj+1, j (j [1, J-1]) J(i) (AARJ) =ADRJ-1, J Department of Automation, Tsinghua University

3. Virtualized Resource Scheduling Average Time Tier j Small VM Average Time Tier j Large VM Service L. Kleinrock, Queueing Systems, Volume 2: Computer Applications. John Wiley and Sons, Inc. , 1976. Average Departure Rate Tier j Departure Average Rate Tier j to Tier Departure j+1 Average Rate Tier j to Tier j - 1 Department of Automation, Tsinghua University

3. Virtualized Resource Scheduling Calculating Response Time AAR 0, 1 AAR 1, 2 A function of AAR 2, 3 A function of AAR 1, 2 AARj-1, j A function of AARj+1, j AARJ-1, J A function of AARj-1, j A function of AAR 1, 2 Department of Automation, Tsinghua University

3. Virtualized Resource Scheduling R(i)(1–P 1)*AST 1 R(i)P 1(1–P 2)*(AST 2+2*AST 1) R(i)P 1 P 2 --- Pj-1(1–Pj)*(ASTj+2*ASTj-1+…+2*AST 1) Department of Automation, Tsinghua University

3. Virtualized Resource Scheduling n Optimization problem Res. Time Cost Min Small Large VM VM Department of Automation, Tsinghua University

4. Experimental Studies n Simulation Toolkits: n n n Real Testbed n n n Matlab Sim. Events IBM X 3950, 16 CPUs 24 GB (Opensuse 11. 1) Apache 2. 0. 55 (1 large VM, 1 small VM) tomcat 5. 5 (2 large VMs, 4 small VMs) My. SQL (1 large VM) Transaction Data: Rubis (an auction site like ebay) Workload. Department Data: of. Web trace from the 1998 Automation, Tsinghua University

4. Experimental Studies Our model suits the workload very well. Our model predicts the response time very well. Department of Automation, Tsinghua University

4. Experimental Studies Utilization based method: Increase or decrease VM based on the utilization of the previous stage. Our method is better than utilization based method. The SLA is satisfied bounded below 10 Sec. The cost is generally less. Department of Automation, Tsinghua University

Thanks n Q&A Department of Automation, Tsinghua University